Tiller

Query your Tiller finance spreadsheet and let an agent write categories and notes back.

Add to Agentman Studio

Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.

Paste it into any MCP client. Setup docs

Resources

What the Tiller connector does

The Tiller MCP server connects your Tiller personal finance spreadsheet — a Google Sheet or Excel workbook that Tiller fills daily with your bank transactions and balances — to Claude, ChatGPT and any MCP-compatible agent. Anthropic's directory lists four tools that read spreadsheets, categories and transactions, and write categories and notes back. Sign-in is OAuth.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on .

Anthropic states this reflects the level of review a connector received, not a security audit.

Tiller tools (4)

  • list_categories
  • list_spreadsheets
  • query_transactions
  • update_transactions

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

  • No application scopes exist. Tiller's authorization server advertises openid, profile, email and offline_access, and its OpenAPI spec requires all four on every financial endpoint. Every one is an identity claim. The RFC 9728 resource descriptor declares no scopes_supported at all. Nothing at the OAuth grant separates reading a decade of transactions from writing to the sheet — the boundary a user would most want falls nowhere in this vocabulary.
  • Tiller calls it beta. Both Tiller blog posts and the help article describe the MCP server as a beta or experimental feature, and the article says of its preset prompts that they have not been thoroughly tested, directing feedback to beta@tiller.com.
  • The write is Google Sheets only. It requires a Direct Fills sheet, and Tiller states Direct Fills is not currently available for Microsoft Excel.
  • 500 transactions per page. Tiller states Claude retrieves 500 at a time and paginates automatically, so wide date ranges are slower.
  • Raw pages come back in sheet order, not by date. Tiller's audit schema states a limited page is the first N sheet rows rather than the newest or oldest.
  • Category names must match exactly. Tiller states your category names are the source of truth, so a question using a synonym finds nothing rather than erroring.
  • Some accounts still need a manual refresh. Tiller states Direct Fills does not remove the requirement to manually refresh institutions that require MFA, so the connector can read a spreadsheet that is silently behind your bank.
  • No undo, and no delete. Nothing in the listing reverts a batch of writes or removes a row. Recovery is your spreadsheet's own version history.
  • The write endpoint's parameters are unpublished. Tiller's OpenAPI spec omits it entirely, so the batch size and the full set of writable fields are unknown to us.
  • ChatGPT is not recommended by the vendor. Tiller reports inconsistent and inaccurate results in early testing and has not validated Gemini.
  • Tiller's privacy policy does not mention the MCP server, AI or LLMs at all. The document the directory links carries a May 2019 revision date, predating the connector and this class of tool entirely. Every privacy commitment about the connector — that Tiller does not store what the AI analyses, that access is explicit and transient, that transactions are not sent to a model automatically — appears only in help-centre and marketing prose, not in the binding policy.
  • Tiller's own AI landing page shipped with unfilled placeholders. On 2026-08-22 it displayed bracketed CMS tokens including a customer-attribution slot and an instruction to insert any error found in its accuracy demonstration, on a page stamped "Verified August 20, 2026". Read its claims as marketing copy in progress rather than as a settled specification.

Frequently asked questions

Can the Tiller connector move money or touch my bank connection?

No. Tiller's own security promise states the platform has read-only access to bank data and that Tiller cannot move, transfer or invest money, create accounts, change passwords or modify permissions at any financial institution. The connector sits above that boundary. Its endpoints read a spreadsheet, and no tool takes a payee, amount or account credential.

What can the Tiller MCP server write to my spreadsheet?

Categories and notes, on Direct Fills sheets only. Tiller's help article states the MCP allows writing changes back to the Transactions sheet for Direct Fills sheets, and that you can use Claude to categorize transactions and apply notes. The documented flow has Claude propose the edits and ask before applying them, so the write is confirmed rather than silent.

What OAuth permissions does the Tiller MCP server request?

Only identity claims. Tiller's authorization server advertised openid, profile, email and offline_access on 2026-08-22, and its OpenAPI spec requires exactly those four for every financial endpoint. None of them separates reading transactions from writing categories, so consent cannot grant read-only access to the connector.

How many tools does the Tiller MCP server have?

Anthropic lists four: list_categories, list_spreadsheets, query_transactions and update_transactions. Tiller publishes no tool names anywhere public — not in its help centre, blog, forum, GitHub, npm or PyPI. Its OpenAPI spec publishes three operations, all GET, all reads, so the write has no machine-readable entry.

Can the Tiller connector see all my Google Sheets?

Not through the connector. It reaches spreadsheets Tiller already services, addressed by a spreadsheet ID that list_spreadsheets returns. Separately, Tiller's Direct Fills authorization asks Google for see, edit, create and delete on all your spreadsheets. Tiller's Advanced Security setting replaces that with a per-file picker, and it is off by default.

How do I know the Tiller connector answered from the right transactions?

Read the audit block. Tiller's query response attaches a record of the executed pipeline with row counts at each step, a truncated flag when a limit cut the tail, and a verdict of ok or warn. Tiller's own schema describes it as letting the caller verify the result matches its intent by reading the plan that produced it.

Why does the Tiller connector miss transactions in a long date range?

Pagination and hidden categories. Tiller states Claude retrieves 500 transactions at a time and paginates through larger sets automatically, so annual analyses take longer. Categories marked hide from reports are excluded by default, and Tiller's schema treats an unresolved hide flag as unknown rather than not hidden, which keeps those rows.

Does the Tiller MCP server work with ChatGPT?

It connects, but Tiller does not recommend it. Tiller states that in early testing with ChatGPT the responses were inconsistent and inaccurate, that it recommends Claude for the best results right now, and that it has not validated Gemini. The endpoint URL is the same for any MCP client, and Tiller does not document setup steps for other tools.

Sources

More Productivity connectors

  • 6sense

    Ask an AI agent about 6sense account insights, buying signals and keyword trends.

  • Adobe Workfront

    Find, create, update and delete Workfront work items from an AI agent.

Browse all MCP connectorsClaude and ChatGPT setup guideBrowse AI skills

Product names and logos belong to their owners. Agentman is not affiliated with a connector's maker unless the page says the connector is built and run by Agentman.

Use it in an agent

Put Tiller to work.

Connect Tiller once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.

Open in Agentman Studio